01 · Seeded state
What does the agent start with?
Policies, claims, payees and reserves for a fictional insurer, Larkspur Mutual, rebuilt from a seed in about 6 ms (p95 8 ms). Every run of a task starts from the same state.
Environments
Simulation environments for AI agents.
An AI agent environment is a working simulation of a business system: a real database, the tools an agent calls, written rules, tasks and graders. Each episode starts from a seeded state and is graded on what the agent changed. AutoStack's live environment is insurance claims; others are built with design partners.
Anatomy
Inside an environment
02 · Tools over MCP
How does the agent act?
Through 15 tools: 6 that read and 9 that write, each backed by the same database the grader checks. Writes change state; that's what gets graded.
03 · Operating manual
What rules apply?
A written claims manual with rule IDs. Every task is written against it, and every failure cites it.
04 · Tasks
What is the agent asked to do?
A public practice set of 15 tasks, each runnable on 3 seeds, including act-or-escalate twins and instructions hidden in customer documents.
05 · Agent actions
What gets recorded?
Every tool call the agent attempts goes into the audit log, including the ones the system rejects.
06 · Grader
How is it scored?
Deterministic checks on the final state and the audit log. Same seed, same result.
Seeded resets
How a seeded reset works
Every episode is rebuilt from a seed, so a failure is something you can reproduce, exactly as it happened. Replaying a recorded episode on the same seed produces the same final state.
run · seed 1
claimLMX-1-002
linePersonal auto · Collision
insurerLarkspur Mutual (fictional)
payment_authority$5,000
deductible$500
tasklk-auth-001
final_state 9c6205397283…
replay · same seed
claimLMX-1-002
linePersonal auto · Collision
insurerLarkspur Mutual (fictional)
payment_authority$5,000
deductible$500
tasklk-auth-001
final_state 9c6205397283…
✓ identical final state
Real episode values: lk-auth-001, seed 1, run 2026-10-08. The replay reproduces the same final state hash.
reset(seed=1) → state built · 6 ms → policies, claims, payees, reserves
replay(seed=1) → same tool calls → same final state ✓
Environment, sandbox or benchmark
Is an environment the same as a sandbox?
A sandbox keeps an agent away from real systems. A benchmark fixes a task set for comparison. An environment does both jobs and grades what changed. What is an RL environment? →
| Property | Sandbox | Benchmark | Environment |
|---|---|---|---|
| Isolated from real systems | Yes | Usually | Yes |
| Realistic data and tools | Empty until you fill it | Fixed | Yes, seeded |
| Tasks with known answers | No | Yes | Yes |
| Grades what the agent changed | No | Sometimes | Yes |
| Use for training rewards | No | Rarely | Yes |
Where agents practice
Live and in development
One is live today: insurance claims. The engine is built for more, and each new one is built with a design partner.
A working P&C claims system for the fictional insurer Larkspur Mutual: policies, coverage, deductibles and limits, reserves, payments, authority limits and fraud referrals. 15 public tasks today.
- 15-task public practice set × 3 seeds
- 15 tools over MCP
- Claims manual with rule IDs
Commercial auto, property and workers' comp, built with a design partner on the same engine.
Building agents for another regulated, high-stakes workflow? Tell us what your agent does and what it must never get wrong.